What is prompt engineering, and do I still need it?

What the Skill Actually Is, before Deciding whether It's Dead

It helps to separate two things that "prompt engineering" has always bundled together.

One is phrasing craft: the wording, the tone-setting, the clever framing, the trick of asking the model to think step by step. The other is specification: stating the goal precisely, supplying the constraints, giving the relevant background, defining the output. Both have lived under the same name, which is why the "is it dead?" debate is so muddled — the answer is different for each.

Phrasing craft is the part that has faded, and it faded for a concrete reason: the models got better at inferring intent. A 2023 model needed coaxing into a sensible answer; a 2026 frontier model mostly understands an ordinarily-worded request and the marginal return on clever phrasing has dropped sharply. Specification has not faded at all. A model that infers your intent perfectly still cannot supply facts it was never given, and "tell it clearly what it's working with" is not a phrasing trick — it's the whole game.

What This means for an Individual

If you are using AI for your own work, prompting skill still pays, and the floor is low: a few habits get you most of the available benefit. The honest allocation has flipped since 2023. Back then, better phrasing often separated unusable output from good output, and time spent on wording paid off. In 2026, most people get acceptable output with ordinary language, and the distance from acceptable to excellent comes from supplying better context and judging the result more sharply, not from a more ingenious way to ask. The skill is still worth having; it is no longer worth most of your attention. State the task and the constraints explicitly. Give an example of the output you want. Say what to leave out, not just what to include. Iterate — read what comes back, see what you under-specified, and say it. That last habit is the durable one, and it is worth noticing that it isn't really a phrasing skill at all. It is the judgement to recognise a weak output and diagnose what was missing, which is discernment wearing a prompt-shaped costume.

That is why the people who stay valuable as models improve are not the ones with the largest prompt libraries. The wording layer of a saved prompt is a frozen guess about what a particular model needed on a particular day, and it decays as models change. What survives is whatever expertise the prompt encodes: the editorial checklist, the legal issue-spotting sequence, the diagnostic questions. But notice that the durable part was never the phrasing. It was the knowledge of what a good answer requires. The transferable skill is exactly that: knowing what a task needs to be done well, what context it takes, what constraints apply, what success looks like — and that skill is indifferent to which model you're using or how good it has become. The prompt is disposable. The judgement that wrote it is not.

What This means for an Organisation

At the individual level, the shift from phrasing to specification is a minor reallocation of attention. At the organisational level it is the whole problem, because specification needs an input that individuals usually supply from their own heads: what the organisation knows, has decided, and is constrained by.

When that input lives only in people's heads, every prompt re-sources it from memory, every person sources it slightly differently, and the same context gets re-typed into a thousand sessions and discarded at the end of each. The work of telling the model what it's working with is real work, and most organisations pay for it over and over without keeping any of it. The individual prompting question — "how do I phrase this?" — has a tidy answer. The organisational version — "where does the context every prompt depends on actually live?" — usually has no answer at all, which is why agents reliably start a task knowing less than the company collectively knows.

This is the point at which prompting stops being the interesting question. If the context an agent needs is authored once and available to every agent and person who needs it, the prompt becomes close to trivial: you are no longer hand-feeding the model what it should already have. The skill doesn't vanish, but its centre of gravity moves from the wording of the request to the quality of the context the request draws on.